4/ We’ve built safety tools like Safe Shield precisely for this
It surfaces warnings for unverified or risky modules such as this one. Including risks like when you're interacting with unrecognized addresses. https://t.co/nR6hCDIlc5
Action for all Safe users:
→ Check Settings → Modules in the app
→ Disable anything unfamiliar immediately
Guide on modules: https://t.co/6vQKjXDV3s
We’re monitoring closely and will update if needed. Your funds in official Safe deployments are safe.
Follow-up to our earlier post:
On May 7, 2026 at 19:50:47 UTC, an attacker attempted to exploit an accounting edge case in the sUSDp savings vault on Ethereum.
The attempt failed. No funds were lost. User funds are safe.
The vector was an ERC4626 inflation / donation-style attack against the sUSDp vault's accounting. In simple terms, the attacker tried to capture an inflated share of the vault's assets by first accumulating a very large percentage of the sUSDp vault supply, then attempting to trigger an inflation event and redeem against an accounting imbalance.
This type of attack only becomes economically viable under very specific conditions, all of which had to hold simultaneously:
1. The attacker must control a very high percentage of the vault supply (>99% of totalSupply on the target chain), so that most of the inflated assets accrue to them rather than being diluted across other holders.
2. The vault must have had a long inactivity window (several weeks with no interaction), so that the dormant accrued yield accumulates and is minted in a single lump during the next interaction.
These conditions were uniquely met on the Ethereum sUSDp vault: a small total supply (~2.7k sUSDp), no accrual for ~40 days (since March 29, 2026), and an attacker who had spent the previous 57 days quietly building a position of 2,672 sUSDp (~99% of the Ethereum supply), split across two independently funded addresses:
- 0xfa8ba3b8ee8691847fbbc9532116ba3f21278b49 (372 sUSDp)
- 0xb6747DF4D24338fD8e1C2b366D23825B35ADc4F3 (2,300 sUSDp)
These two positions were funded from sources not directly linked on-chain to the attacker's operating address, which is the address used to deploy the exploit contracts and submit the attack transactions: 0x6375c37761a26fac9b201af9f527caa9627aea11
Estimated impact. Based on our auditors' analysis of the attack transaction, had the exploit succeeded, the attacker would have ended up with over 1.52M USDp. The mechanism would have unfolded as follows:
1. The attacker contract first transfers the sUSDp shares from the two preparation wallets, giving it 99.88% of the vault's totalSupply.
2. The contract then deposits 223.83M USDp into the savings vault.
3. The dormant fixed-rate interest accrual on the inflated balance mints an additional 1.525M USDp into the vault, bringing total assets to 225.355M USDp.
4. The contract redeems its 99.88% share, walking away with approximately 1.52M USDp.
Forensic timeline (UTC):
- A few hours before the attack: the attacker created a Safe (https://t.co/mudd1LXwuq). We are not certain of the intent, but our working hypothesis is that this was an attempt to disguise the on-chain pattern from monitoring systems such as Hypernative.
- 19:50:47: The attacker initiated his first attempt by deploying an exploit contract (https://t.co/v3932U5aoK). The attack transaction was submitted shortly thereafter and reverted on-chain. Based on our current analysis (still being confirmed), the revert was caused either by insufficient gas to cover the LayerZero fee for the bridge leg of the exploit, or by attempting to bridge an amount exceeding our daily cross-chain limit (2.5M USDp). Either way, the exploit did not execute.
- ~30 to 45 seconds later: @HypernativeLabs' live monitoring detected the pattern, and the relevant contracts were paused within a few blocks of the failed transaction.
- 20:52:23: The attacker deployed a second exploit contract (https://t.co/NN3Fwa5krh) and submitted a second attack transaction. This one reverted because the contracts were already paused.
- After the second failure: the attacker forwarded his remaining ETH through several intermediary wallets in sequence before sending it to what we believe to be a KuCoin address (https://t.co/kJB95cSYEo). This appears to be another attempt to obscure the trail.
Why the attack failed: defense in depth. The attempt was stopped by three independent layers, any one of which would have been sufficient on its own.
1. The attacker's own attack transaction failed on-chain. Based on our preliminary analysis (still being confirmed), the failure was caused either by insufficient gas for the LayerZero fee, or by hitting the 2.5M USDp daily cross-chain bridge limit. The exploit therefore never executed.
2. @HypernativeLabs' live monitoring detected the pattern shortly after, and the relevant contracts were paused within a few blocks (~30 to 45 seconds) of the failed transaction. By the time the attacker re-attempted at 20:52:23 UTC, the contracts were already paused.
3. Bridge finality would have trapped the funds. Even in the counterfactual where the first transaction had succeeded and the bridge had been triggered, LayerZero finality from Ethereum to Base takes ~4 minutes. The pause of lz-USDp (within ~30 to 45 seconds) would have caught the in-flight bridge message well before delivery on Base, and the exploited USDp would have been frozen in the bridge.
In other words, the attacker needed his transaction to land, the contracts to stay live for the next 4+ minutes, and the bridge to deliver. All three failed.
Status of the attacker's position. With sUSDp paused, the attacker's 2,672 sUSDp are frozen. He cannot redeem, transfer, or otherwise access these funds. The address at the end of the on-chain trail, which we believe to be a KuCoin address, is subject to KYC, so the attacker is potentially identifiable.
Why this is bounded. The same conditions that made sUSDp on Ethereum exploitable (small supply, multi-week inactivity, adversary holding ~all of it) do not currently exist on our other deployments: Base, HyperEVM, and Avalanche. The attack vector requires both a supply concentration and a dormancy window that are easy to monitor and prevent.
Current status:
- No funds were lost
- User funds are safe
- USDp remains transferable
- The Savings Module (sUSDp) remains paused
- The Bridging Module (lz-USDp) and Parallelizer Module remain paused for now as a precaution
Based on the auditors' recommendations, the Bridging Module and Parallelizer Module will be unpaused once the DAO multisig signers tighten the bridging, minting, and burning limits.
The Savings Module will remain paused until the vault accounting patch is finalized and audited. As an immediate operational measure, on the auditors' recommendation we have already issued a setRate() call across all four deployments (Ethereum, Base, HyperEVM, Avalanche) to reset the accrual state and clear any accumulated dormant lump. Two independent audits will be performed on the patched contracts, by @bailsecurity and @cyfrin, before redeployment. The module will only resume once we and our partners are fully confident the vector is closed.
This is exactly why we operate with multiple layers of security: audits, bug bounty, live monitoring, emergency pause procedures, backup domains, and additional operational safeguards. Today, every layer was tested and held.
A full post-mortem will follow in the coming days with the complete technical write-up, timeline, remediation steps, and next actions.
Earlier today, we detected an attempted exploit against Parallel.
@HypernativeLabs’s live monitoring triggered in time, and contracts were paused. The attempt failed. No funds lost; user funds are safe.
A huge thank you to @fraxfinance, @bailsecurity, @cyfrin, @merkl_xyz, and @GamiLabs for their reactivity, support, and help in handling the situation quickly.
An update will be provided in the coming hours, giving more details.
Full post-mortem in the coming days.
Our friends at @tulipacapital and @HypernativeLabs alerted us about something strange happening on chain with @yieldbasis pool positions moving.
We or any other of our users are totally UNAFFECTED but smart contracts of a project which was experimenting with building on top was.
They were notified immediately. We love our users and try to help them to stay safe whenever we can!
“We’re going to partner with Mr. beast on new financial and content primitives”
Translation ->
“we are going to shill ETH worldwide to everyone under 35 so hard it’ll make your teeth hurt”
What crypto people get wrong is the “3% of txn fees” is for CREDIT card purchases (aka unsecured consumer loans). Unsecured loans doesn’t exist in crypto today (or certain not like traditional cc loans).
Crypto card purchases act like debit card purchases. Debit purchases charge merchants more like ~20 cents per txn.
Stablecoins don’t provide improvements to legacy rails from a cost standpoint vs debit card purchases..
there’s a common view that card networks (visa, ma, etc) should get disrupted by new rails such as as stablecoins because they charge merchants an absurd 3% fee.
the reality is card networks keep a v small portion of this fee. assume a typical $3 fee on a $100 purchase, about
- $1.8 goes to the consumer (as card rewards)
- $0.45 goes to the issuing bank (the consumer’s bank)
- $0.6 goes to the acquiring bank (bank of the merchant)
- $0.15 goes to visa/ma
this is one of the most remarkable examples of incentive alignment:
- consumers keep a lion share of it as incentives to continue to use the card
- issuing bank is incentivized to acquire customers for the network
- acquirer is incentivized to acquire merchants for the network
- visa/ma keep a tiny portion but they get to scale the network at virtually 0 marginal cost
- merchants pay for all of this as they have the least bargaining power
easily one of the strongest instances of network effect we’ve seen in the history of businesses
On the surface this seems like the right bet until you realize klarna, affirm, et al pay the merchants upfront (minus merchant fee) and aren’t really in the black until the final payment is made by the borrower
The "buy now pay later" epidemic is spiraling out of control
I've been in Chile for the last week and with every single purchase (groceries, dinner, coffee)
They ask if I want to pay for everything up front, or break it up into 6-12 interest-free payments
Turns out people are basically financing every single purchase they make - instead of paying $50 for groceries, they'll pay $5/mo for 10 months
Longing Klarna stock at $30 and betting on this trend continuing in the USA feels like free money
My second favorite college football team, Vanderbilt, may be getting a raw deal from the CFP committee—like, maybe it should be ranked as high as 6th, rather than 14th. Below is my “Common Opponent” analysis.
Common Opponent Analysis: Why the CFP may be Severely Underrating Vanderbilt
Tim Groseclose
Dec. 6, 2025
In this document I use what I call Common Opponent Analysis to compare Vanderbilt with other teams that are ranked among the top 25 in the College Football Playoff ranking.
As the analysis shows, the current CFP ranking seems to place Vanderbilt significantly worse than the rank it may deserve.
Defining the Method, an Example
To illustrate the method, let us first consider one team, Oklahoma, as an example. Oklahoma and Vanderbilt had seven common opponents: Texas, Auburn, South Carolina, Tennessee, Alabama, Missouri, and LSU.
Vanderbilt lost to Texas by 3 points. Importantly, however, Texas was the home team in the game. Various analyses have estimated the home-field advantage to be approximately 3 points. Thus, if we discount home-field advantage, the game suggests that—on a neutral field—Vanderbilt and Texas would be evenly matched. That is, Texas is 0 points better than Vanderbilt.
Meanwhile, Oklahoma lost to Texas 23-6. The game was played on a neutral field. Thus, the game suggests that Texas is 17 points better than Oklahoma.
If we combine the two games---Oklahoma v. Texas and Vanderbilt v. Texas---they suggest that Vanderbilt is 17 points better than Oklahoma.
If we consider the other six common opponents, the games imply other amounts by which Vanderbilt is better or worse than Oklahoma. Indeed, some of the games suggest that Vanderbilt is worse than Oklahoma. Using the above method on the other six opponents respectively gives the following amounts by which Vanderbilt is better than Oklahoma. (A negative number indicates that---using the two games of the particular common opponent---Oklahoma is better than Vanderbilt.). Auburn, 0; South Carolina, 5; Tennessee, 15; Alabama -18; Missouri, -2; LSU 3.
Finally, I compute the average of the above seven numbers. It equals 2.86. (Specifically [17+0+5+15-18-2+3]/7 = 20/7.). Thus, the Common Opponent Method implies that Vanderbilt is 2.86 points better than Oklahoma.
Results: How the Common Opponent Method Rates Vanderbilt Against Other Teams in the CFP Top 25
Not counting Vanderbilt, there are 24 teams in the CFP top 25. Of these 24 teams, ten had at least one common opponent with Vanderbilt. I list these teams, their current ranking in the CFP, the common opponents that the team shared with Vanderbilt, and how the Common Opponent Method Rates the team against Vanderbilt.
1. Ohio State. Vanderbilt and Ohio State had one common opponent, Texas. The Common Opponent Method rates Ohio State as 4 points better than Vanderbilt.
3. Georgia. Vanderbilt and Georgia had five common opponents: Tennessee, Alabama, Kentucky, Auburn, and Texas. The Common Opponent Method rates Georgia as 3.8 points better than Vanderbilt.
6. Ole Miss. Vanderbilt and Ole Miss had three common opponents: Kentucky, LSU, and South Carolina. The Common Opponent Method rates Vanderbilt as 10.33 points better than Ole Miss.
7. Texas A&M. Vanderbilt and Texas A&M had six common opponents: Utah State, Auburn, LSU, Missouri, South Carolina, and Texas. The Common Opponent Method rates Texas A&M as 1.33 points better than Vanderbilt.
8. Oklahoma. Vanderbilt and Oklahoma had seven common opponents: Texas, Auburn, South Carolina, Tennessee, Alabama, Missouri, and LSU. The Common Opponent Method rates Vanderbilt as 2.86 points better than Oklahoma.
9. Alabama. Vanderbilt and Alabama had five common opponents: Missouri, Tennessee, South Carolina, LSU, and Auburn. The Common Opponent Method rates Vanderbilt as 3 points better than Alabama.
(It should be noted that Vanderbilt and Alabama actually played each other. In that game Alabama won by 16 points. Alabama, however, was the home team. Thus, if we discount the home-field advantage, the score suggests that Alabama is 13 points better than Vanderbilt. It should also be noted that Alabama scored a “junk” touchdown at the end of the game. That is, Alabama, trying to run out the clock, did a run play near the end of the game. Vanderbilt, possibly trying to strip the ball from the runner rather than making a safe tackle, allowed the runner to score a lucky touchdown. The play illustrates (i) that Alabama more accurately is only about 6 points better than Vanderbilt and (ii) in any game there are many random incidents that can cause the score to misrepresent the true degree by which one team is better than the other. For this reason---and because the Common Opponent Method often involves a larger sample of games---the Common Opponent method might actually be a better method than using a head-to-head matchup of the two teams.)
12. Miami. Vanderbilt and Miami had one common opponent, Virginia Tech. The Common Opponent Method rates Vanderbilt as 7 points better than Miami.
13. Texas. Vanderbilt and Texas had one common opponent, Kentucky. The Common Opponent Method rates Vanderbilt as 19 points better than Texas.
(It should be noted that Vanderbilt and Texas actually played each other. In that game Texas won by 3 points. Texas, however, was the home team. If we discount the home-field advantage, the game suggests that Texas and Vanderbilt are equally matched. It should also be noted that at the end of the game Vanderbilt recovered an onside kick, however did so just barely out of bounds, which gave the ball to Texas. Importantly, if the ball had bounced a few millimeters differently, then Vanderbilt would have recovered the ball in bounds and would have had another chance to score. The incident is another example of the random factors in a football game, thus suggesting the importance of a larger sample size, as the Common Opponent Method uses, rather than a sample of one, as a head-to-head matchup uses.)
17. Virginia. Vanderbilt and Virginia had one common opponent, Virginia Tech. The Common Opponent Method rates Vanderbilt as 10 points better than Virginia.
22. Georgia Tech. Vanderbilt and Georgia Tech had one common opponent, Virginia Tech. The Common Opponent Method rates Vanderbilt as 15 points better than Georgia Tech.
Additional Notes
Notre Dame is number 10 in the CFP rankings. Although Notre Dame and Vanderbilt had no common opponents, Notre Dame played Texas A&M and Miami. After discounting home-field advantage, the games suggest that Notre Dame is even with Miami, while Texas A&M is 2 points better than Notre Dame. As I note above, the Common Opponent Method rates Vanderbilt as 7 points better than Miami. If Notre Dame is even with Miami, this suggests that Vanderbilt is 7 points better than Notre Dame. As I note above, the Common Opponent Method rates Texas A&M as 1.33 points better than Vanderbilt. If Texas A&M is 2 points better than Notre Dame, this suggests that Vanderbilt is .67 points better than Notre Dame. The average of the above two numbers is 3.84 ( = [7+.67]/2), thus suggesting that Vanderbilt is 3.84 points better than Notre Dame.
Michigan is number 19 in the CFP rankings. Although Michigan and Vanderbilt had no common opponents, Michigan played Oklahoma and Ohio State. After discounting home-field advantage, the games suggest that Oklahoma is 8 points better than Michigan, and Ohio State is 21 points better than Michigan. As I note above, the Common Opponent Method rates Vanderbilt as 2.86 points better than Oklahoma. If Oklahoma is 8 points better than Michigan, this suggests that Vanderbilt is 10.86 (=8+2.86) points better than Michigan. As I note above, the Common Opponent Method rates Ohio State as 4 points better than Vanderbilt. If Ohio State is 21 points better than Michigan, this suggests that Vanderbilt is 17 points better than Michigan. The average of the above two numbers is 13.93 ( = [10.86+17]/2), thus suggesting that Vanderbilt is about 14 points better than Michigan.
Tulane is number 20 in the CFP rankings. Although Tulane and Vanderbilt had no common opponents, Tulane played Ole Miss. After discounting home-field advantage, the game suggests that Ole Miss is 32 points better than Tulane. As I note above, the Common Opponent Method rates Vanderbilt as 10.33 points better than Ole Miss. If Ole Miss is 32 points better than Tulane, this suggests that Vanderbilt is 42.33 (=32+10.33) points better than Tulane.
Georgia Tech is number 22 in the CFP rankings. Although Georgia Tech and Vanderbilt had no common opponents, Georgia Tech played Georgia. After discounting home-field advantage, the game suggests that Georgia is 7 points better than Georgia Tech. As I note above, the Common Opponent Method rates Georgia as 3.8 points better than Vanderbilt. The difference in the two numbers suggests that Vanderbilt is 3.2 points better than Georgia Tech.
Real talk. I can't recommend the @HypernativeLabs stack enough.
They make it incredibly easy to spin up on chain agents that can do block by block evaluations (Events, Function Calls, Etc.), API Integration (For Web2 Calls), Calculations and free form Python.
Safe just got safer.
We're partnering with @SafeLabs_ to bring enterprise-grade transaction security directly into the world's leading multisig.
❌No separate tools.
❌No extra logins.
✅Just institutional-grade protection embedded exactly where you need it.
Hypernative can now screen every Safe transaction for 300+ risk types before execution. Real-time, zero-day threat detection, custom policies, and onchain enforcement. All native to your Safe workflow.
$65B in assets. Zero compromise on security or speed.
📰Read the announcement: https://t.co/5bVB6LOvH8
🛡️Safe Shield: https://t.co/wmoiSRPHfH
🔗Book a demo: https://t.co/lDRpKjcrXi